Steering two orthogonal concept neurons makes a model compose "gnome"
Sauers_ · x · 2026-10-27
@Sauers shares an interpretability demo: steering neurons for "GNOME desktop (I'm a screen)" and "small elf (creature)" made the model compose them into "gnome (creature)". The quoted thread floats testing whether "Blue" + "Hedgehog" yields "I am Sonic".
More from Fun
- User spots eerie moment: device starts running on its own with no user action — generativist · 2026-10-27
- 'We're not bad at forecasting—I'm bad at QA-ing my posts,' dlwh wraps up — dlwh · 2026-10-27
- Dev shows striking SD vs HD game texture upgrade alongside a level editor — maxleiter · 2026-10-27
- AI-generated movie-grade clip goes viral, blurring the line with real cinema — cneuralnetwork · 2026-10-27
- Drinking under the JD dog: Beijing's satellite-factory bar next to Moonshot's faux temple — teortaxesTex · 2026-10-27
- Creator's 2566-rated champion bot dethroned on Arena Ladder, training a comeback version — tristanbob · 2026-10-27